Hierarchical Bayesian Personalized Recommendation: A Case Study and Beyond

Yan Yan; Zhexuan Xu; Zitao Liu

REVIEW 5 major objections 5 minor 34 references

Hierarchical Bayesian Personalized Recommendation: A Case Study and Beyond

T0 review · 5 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a hierarchical Bayesian model in which each brand inherits exactly one latent style can beat established recommender baselines on click prediction.

desk verdict Plausible model, but the headline empirical gains may rest on test-set leakage in the popularity feature; the paper never says how POP is computed per fold. read the letter →

arxiv 1908.07371 v1 pith:EJVZ4LK2 submitted 2019-08-20 cs.LG cs.IRstat.ML

Zitao Liu , Zhexuan Xu , Yan Yan This is my paper · ORCID

classification cs.LGcs.IRstat.ML

keywords hierarchicalBayesianmodelsrecommendersystemsvariationalinferencelatentstylemodelingclickpredictioncollaborativefilteringpersonalizedrankinge-commerce

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes HBayes, a hierarchical Bayesian model for personalized recommendation that learns both the item hierarchy and the latent factors attached to it. The key claim is that explicitly modeling each brand as drawing one hidden 'style,' with all of the brand's products inheriting that style, lets a recommender beat established baselines on click prediction. The paper tests this on two real-world datasets, an e-commerce apparel catalog and a music listening dataset, and reports gains over the comparison models in precision, recall, and normalized discounted cumulative gain. A sympathetic reading is that the model offers a transparent, interpretable alternative to deep black-box recommenders for settings where products naturally live in a hierarchy.

What carries the argument

The load-bearing object is the HBayes generative process plus its variational inference routine. Each brand draws a one-hot style assignment from a Dirichlet-multinomial prior; its brand vector is then Gaussian centered at the chosen style vector, so the hierarchy enters the model as conditional independence rather than as extra features. Inference uses the mean-field approximation with a sigmoid lower bound that restores conjugacy between the logistic click likelihood and the Gaussian priors, giving closed-form coordinate-ascent updates for every factor. That mechanism is what lets the model learn style clusters and user and brand vectors jointly.

What would settle it

Run HBayes on a catalog whose brands are known to span multiple style categories and inspect each brand's posterior over styles. If many brands show diffuse, split assignments while a version of the model that allows a per-brand style mixture matches or beats HBayes on NDCG, then the single-assignment assumption, not the hierarchy itself, is carrying the result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a generative story—user vectors drawn from Gaussian priors, brands assigned to one of several latent styles by a multinomial draw, style vectors drawn from a shared Gaussian, and click probability logistic in the inner product of item features with brand plus user vectors—can be learned by a mean-field variational Bayes algorithm and yields top-K recommendations that beat HPF, HSR, SVD++, factorization machines, co-clustering, and LambdaMART on both an apparel dataset and a music dataset. The learned style assignments are also presented as interpretable: projections of the learned brand and style vectors show clusters that correspond to recognizable garment styles. The paper further claims the variational procedure converges in few iterations, making the model practical despite being fully Bayesian.

Load-bearing premise

The model assumes every brand has exactly one hidden style and all of that brand's products inherit it; if real brands span multiple styles, the learned hierarchy misses structure and the reported gains may not generalize.

Editorial extensions

If this is right

If HBayes is right, recommenders can exploit an explicit product hierarchy without flattening it into a huge feature space or handing the hierarchy to a deep network to rediscover.
The reported NDCG gains grow with list length on the apparel data, suggesting the main benefit is better ranking of items deeper in the recommendation list, not just top-5 accuracy.
Because the posterior gives distributions over user, brand, and style vectors, a deployed system could carry uncertainty into downstream decisions such as exploration or explanation.
The framework's success on two domains implies the same generative pattern may transfer to other hierarchically organized catalogs such as news, video, or grocery.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

I would not extend the single-style-per-brand assumption to general catalogs: the generative process draws exactly one style per brand, so brands that genuinely span several styles are forced into one cluster; a natural test is whether allowing a per-brand mixture over styles preserves or improves the reported gains.
The paper optimizes a pointwise logistic likelihood; a pairwise or listwise objective, as in Bayesian personalized ranking, might combine even better with the same hierarchical prior, since the evaluation metrics are ranking-based.
One can read the learned style assignments as a cheap taxonomy-discovery tool: applying HBayes to a catalog without labels should recover meaningful product groupings, which could be verified by human raters.
The convergence speed claim suggests the variational updates might scale to much larger catalogs, but the experiments cap at 895 users and roughly 380,000 events, so the practical ceiling is untested.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Before you cite this paper, know this: the model is a workmanlike combination of known parts, but the empirical headline may be an artifact of test-set leakage in the POP feature, and the paper never addresses it.

What is new: a Dirichlet-multinomial 'style' layer between brands and items, Gaussian latent factors, and a logistic click model with Jaakkola-Jordan variational inference. The generative story is coherent, and the updates in Section III-C are internally consistent. The empirical work covers two datasets, one proprietary e-commerce set and Last.fm, and the authors report gains over several baselines. That part is fine as a case study.

The soft spots are not minor. In Section IV-B1, POP is defined as a ratio where the denominator is the sum over 'all products in the dataset.' Under 5-fold CV, that means the item's feature vector uses a global popularity statistic computed over the entire dataset, including test events. The authors never state that POP is recomputed per training fold. If it is not, the strong NDCG margins over HPF (10-15% on apparel) could simply be popularity leakage, because HPF does not consume content features at all. This is the first thing a referee should check. There are also no error bars or significance tests, and Table II has an obvious typo (NMF NDCG@10 = 0.0156, which is clearly off by a decimal). The single-style-per-brand assumption is a real limitation: brands are assigned one latent style via a multinomial draw, so the model is not learning a hierarchy so much as a one-hot assignment.

The math itself looks sound; the variational inference is standard. But the evaluation section is under-specified: no hyperparameter settings (S, d, gamma priors), no split details, no negative sampling description. The tiny URL for code may or may not work; regardless, we cannot verify the pipeline.

Who gets value from this: researchers combining Bayesian latent factor models with hierarchical item taxonomies. It is an incremental contribution, not a foundation-shaker. A serious reviewer should ask for a leakage-safe re-evaluation with per-fold POP and per-fold category weights, plus a table of hyperparameters. If the gains survive, the paper is publishable at a decent venue; if not, it is a useful cautionary tale about popularity features in recommender evaluation.

My recommendation: send it out for review, but with these flags. It deserves a serious referee, and the questions the referee will ask are exactly the ones the authors should have answered.

Referee Report

5 major / 5 minor

Summary. This manuscript introduces HBayes, a hierarchical Bayesian framework for personalized recommendation. The generative model associates each brand with a latent 'style' drawn from a multinomial distribution, places Gaussian priors on user, style, and brand factors, and models click events through a logistic likelihood. The authors derive a mean-field variational Bayes algorithm with a Jaakkola-Jordan sigmoid bound, and evaluate the model on an e-commerce apparel dataset and a Last.fm/FMA music dataset, reporting precision, recall, and NDCG against seven baselines.

Significance. If the empirical evaluation is sound, HBayes is a useful contribution: it provides an interpretable, non-deep alternative that explicitly models latent styles at the brand level, and the authors make code publicly available. The variational updates are internally consistent, and the prediction formula uses posterior expectations from training-fold learned parameters, so the central performance claim is not circular. However, the evaluation section leaves several load-bearing protocol details unspecified (data split, negative sampling, feature leakage, hyperparameters), so the significance of the reported gains is currently conditional.

major comments (5)

[Section III-A, Eq. (3), Eq. (11)] The generative process is ambiguous about whether the style-proportion variable θ is global or brand-specific. Step 3 states 'Draw a style proportion distribution θ for each brand i,' which suggests θ_i, but the log-joint in Eq. (3) contains only one log p(θ|γ) term, and the update in Eq. (11) updates a single global γ as γ_j = γ_j + ∑_i μ_i,j. This is only coherent if θ is global. Please clarify the intended model and align Step 3, the graphical model, Eq. (3), and Eq. (11); as written, the model definition is internally inconsistent.
[Section IV-A] The evaluation protocol is under-specified. The paper does not state whether the 5-fold split is by user or by event, how negative examples were constructed for the apparel data (or whether the test set is a mixture of positives and negatives), or what candidate set is ranked when generating 'top K samples.' Without these details, the precision, recall, and NDCG numbers in Figures 2-3 and Tables II-III are not reproducible and their meaning is unclear.
[Section IV-B1] The POP feature is defined as n_{x_i}/N_x over 'all products in the dataset' with no statement that it is recomputed from each training fold only. If this aggregate is computed on the full data before the 5-fold split, test-period popularity information enters the training features; because HPF 'does not leverage the entity content feature' while HBayes consumes POP, the Table II margins over HPF could reflect popularity leakage rather than the hierarchical model. Please recompute all content features within each fold, and report the comparison both with and without per-fold feature computation.
[Tables II and III] No error bars, standard deviations, or significance tests are reported for any metric. The claim that HBayes outperforms HPF is based on margins as small as 2.7% (Last.fm NDCG@50, Table III), and without fold-level variance it is not possible to tell whether the differences are meaningful.
[Section III-C2 and Section IV] The manuscript never reports the values of key hyperparameters: the number of styles S, the latent dimension d, the Gamma shape/rate parameters α and β, or the initialization of the variational parameters. Figure 4 shows four learned style clusters, but S is not stated. These omissions prevent reproduction of the experiments even with the public code.

minor comments (5)

[Table II] The NMF NDCG@10 entry, 0.0156, is inconsistent with the surrounding values (0.1249 at K=5 and 0.2272 at K=25) and looks like a typo; please correct it.
[Section IV-B] The feature dimensions do not add up: 20 (popularity+category) + 50 (title) + 50 (property) = 120, not the stated 140.
[Section IV-C] The music experiment does not describe how the FMA precomputed features are used or which features are fed to HBayes and the baselines.
[Algorithm 1] No convergence criterion for the variational updates is given; please state the stopping condition and the number of iterations used in practice.
[Section I] The phrase 'contributions in four folds' should be 'four facets' or 'fourfold'; this is a typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: HBayes is a self-contained generative model with held-out empirical evaluation.

full rationale

The derivation chain is self-contained. The generative process in Section III-A defines priors and likelihood; the variational updates in eqs. (9)-(19) optimize the free-energy lower bound; and the prediction rule in Section III-D is the standard posterior predictive expectation under the fitted variational posterior, not a fitted quantity. The empirical comparison uses 5-fold cross-validation on held-out test events for both datasets, so the headline NDCG, precision, and recall numbers are measured rather than implied by the model definition. I find no load-bearing self-citation: the cited variational bounds ([19], [20]) and baselines are external, and the authors do not cite their own prior work as the basis for the model. The potential aggregate-feature leakage in the POP_i definition in Section IV-B1 is a correctness and validity concern, but it is not circularity under the defined patterns because the reported metrics are still computed on held-out labels and no fitted parameter is renamed as a prediction. The qualitative style-cluster interpretation in Section IV-B3 is a self-assessment of the model's own latent variables, but it is not a derivation of a claimed result and does not reduce the central claim to its inputs. Overall, no circular step rises to the level of a quoteable reduction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim that HBayes outperforms baselines rests on the model's generative assumptions, the variational approximation, and several unstated hyperparameters. The most consequential choices are the number of latent styles and the latent dimension, neither of which is specified in the paper.

free parameters (4)

S
Number of latent style clusters. A core capacity parameter of the model, never stated in the paper.
d
Dimension of user, brand, and style latent vectors. Not reported.
alpha, beta
Gamma prior hyperparameters for precision variables. Not specified in the paper.
gamma = 1/S
Dirichlet prior concentration, initialized to 1/S in Section III-B.

assumptions (5)

domain assumption Each brand i has a single style assignment z_i drawn from Mult(theta).
Section III-A Step 4.1. Restricts each brand to one latent style, a strong claim about the data generating process.
domain assumption Mean-field factorization q(Z,Theta) = product of independent marginals.
Section III-C eq (6). Standard approximation but may be inaccurate when posterior correlations are strong.
standard math Jaakkola-Jordan lower bound for the sigmoid is tight enough for inference.
Section III-C1. The bound is a known variational technique, but its accuracy for this model is not assessed.
domain assumption Click labels are conditionally independent given user and brand factors.
Section III-A generative process Step 5. Ignores temporal or contextual dependencies.
domain assumption The brand-to-item hierarchy is known and fixed.
Only one level of hierarchy is used; the paper does not infer tree structure.

invented entities (1)

Latent style variables S_j
purpose: Represent hidden stylistic categories that connect brands to products
The styles are inferred from the same data used for evaluation. Visual inspection via t-SNE is not a falsifiable external test.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical Bayesian Personalized Recommendation: A Case Study and Beyond." pith.science (2026). https://pith.science/paper/EJVZ4LK2

@misc{pith2026190807371,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Bayesian Personalized Recommendation: A Case Study and Beyond},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EJVZ4LK2}},
  note         = {Machine review of arXiv:1908.07371}
}

read the original abstract

Items in modern recommender systems are often organized in hierarchical structures. These hierarchical structures and the data within them provide valuable information for building personalized recommendation systems. In this paper, we propose a general hierarchical Bayesian learning framework, i.e., \emph{HBayes}, to learn both the structures and associated latent factors. Furthermore, we develop a variational inference algorithm that is able to learn model parameters with fast empirical convergence rate. The proposed HBayes is evaluated on two real-world datasets from different domains. The results demonstrate the benefits of our approach on item recommendation tasks, and show that it can outperform the state-of-the-art models in terms of precision, recall, and normalized discounted cumulative gain. To encourage the reproducible results, we make our code public on a git repo: \url{https://tinyurl.com/ycruhk4t}.

Figures

Figures reproduced from arXiv: 1908.07371 by the authors.

**Figure 2.** Precision@K on Apparel. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 recall@5 recall@10 recall@25 recall@50 Recall [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 5.** Apparel examples in each style clusters. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 6.** Figure 6: Precision@K on Music. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 recall@5 recall@10 recall@25 recall@50 Recall [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

[1]

Item-based collabo- rative ﬁltering recommendation algorithms,

B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-based collabo- rative ﬁltering recommendation algorithms,” in Proceedings of the 10th International Conference on World Wide Web , ser. WWW ’01. New York, NY , USA: ACM, 2001, pp. 285–295

work page 2001
[2]

A survey of collaborative ﬁltering techniques,

X. Su and T. M. Khoshgoftaar, “A survey of collaborative ﬁltering techniques,” Adv. in Artif. Intell. , vol. 2009, pp. 4:2–4:2, Jan. 2009

work page 2009
[3]

Factorizing personalized markov chains for next-basket recommendation,

S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme, “Factorizing personalized markov chains for next-basket recommendation,” in WWW, 2010, pp. 811–820

work page 2010
[4]

P. Lops, M. de Gemmis, and G. Semeraro, Content-based Recommender Systems: State of the Art and Trends , 2011, p. 73

work page 2011
[5]

How do users describe their information need: Query recommendation based on snippet click model,

Y . Liu, J. Miao, M. Zhang, S. Ma, and L. Ru, “How do users describe their information need: Query recommendation based on snippet click model,” Expert Systems with Applications , vol. 38, no. 11, pp. 13 847 – 13 856, 2011

work page 2011
[6]

Who, where, when, and what: A nonparametric bayesian approach to context-aware recommendation and search for twitter users,

Q. Yuan, G. Cong, K. Zhao, Z. Ma, and A. Sun, “Who, where, when, and what: A nonparametric bayesian approach to context-aware recommendation and search for twitter users,” ACM Trans. Inf. Syst. , vol. 33, no. 1, pp. 2:1–2:33, Feb. 2015

work page 2015
[7]

Social recommendation: a review,

J. Tang, X. Hu, and H. Liu, “Social recommendation: a review,” Social Network Analysis and Mining , vol. 3, no. 4, pp. 1113–1133, 1 2013

work page 2013
[8]

Collaborative ﬁltering with temporal dynamics,

Y . Koren, “Collaborative ﬁltering with temporal dynamics,” Commun. ACM, vol. 53, no. 4, pp. 89–97, Apr. 2010

work page 2010

Show all 34 references

[9]

Temporal dynamics of user interests in tagging systems,

D. Yin, L. Hong, Z. Xue, and B. D. Davison, “Temporal dynamics of user interests in tagging systems,” in Proceedings of the Twenty-Fifth AAAI Conference on Artiﬁcial Intelligence, ser. AAAI’11. AAAI Press, 2011, pp. 1279–1285

2011
[10]

Bpr: Bayesian personalized ranking from implicit feedback,

S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized ranking from implicit feedback,” in UAI. AUAI Press, 2009, pp. 452–461

2009
[11]

Groove radio: A bayesian hierarchical model for personalized playlist generation,

S. Ben-Elazar, G. Lavee, N. Koenigstein, O. Barkan, H. Berezin, U. Paquet, and T. Zaccai, “Groove radio: A bayesian hierarchical model for personalized playlist generation,” in WSDM. ACM, 2017, pp. 445– 453

2017
[12]

Efﬁcient bayesian hierarchical user modeling for recommendation system,

Y . Zhang and J. Koren, “Efﬁcient bayesian hierarchical user modeling for recommendation system,” in SIGIR. ACM, 2007, pp. 47–54

2007
[13]

Personalized recommendation in social tagging systems using hierarchical clustering,

A. Shepitsen, J. Gemmell, B. Mobasher, and R. Burke, “Personalized recommendation in social tagging systems using hierarchical clustering,” in Recsys. ACM, 2008, pp. 259–266

2008
[14]

Exploring hierarchical structures for recommender systems,

S. Wang, J. Tang, Y . Wang, and H. Liu, “Exploring hierarchical structures for recommender systems,” TKDE, 2018

2018
[15]

Multiple hypergeometric functions: Probabilistic in- terpretations and statistical uses,

J. M. Dickey, “Multiple hypergeometric functions: Probabilistic in- terpretations and statistical uses,” Journal of the American Statistical Association, vol. 78, no. 383, pp. 628–637, 1983

1983
[16]

Approximate bayesian inference for latent gaussian models by using integrated nested laplace approx- imations,

H. Rue, S. Martino, and N. Chopin, “Approximate bayesian inference for latent gaussian models by using integrated nested laplace approx- imations,” Journal of the royal statistical society: Series b (statistical methodology), vol. 71, no. 2, pp. 319–392, 2009

2009
[17]

Pattern recognition and machine learning (information science and statistics) springer-verlag new york,

C. M. Bishop, “Pattern recognition and machine learning (information science and statistics) springer-verlag new york,” Inc. Secaucus, NJ, USA, 2006

2006
[18]

Latent dirichlet allocation,

D. M. Blei, A. Y . Ng, and M. I. Jordan, “Latent dirichlet allocation,” Journal of machine Learning research , vol. 3, no. Jan, pp. 993–1022, 2003

2003
[19]

A variational approach to bayesian logistic regression models and their extensions,

T. Jaakkola and M. Jordan, “A variational approach to bayesian logistic regression models and their extensions,” in Sixth International Workshop on Artiﬁcial Intelligence and Statistics , vol. 82, 1997, p. 4

1997
[20]

An introduction to variational methods for graphical models,

M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul, “An introduction to variational methods for graphical models,” Machine learning, vol. 37, no. 2, pp. 183–233, 1999

1999
[21]

Exploring implicit hierarchical structures for recommender systems

S. Wang, J. Tang, Y . Wang, and H. Liu, “Exploring implicit hierarchical structures for recommender systems.” in IJCAI, 2015, pp. 1813–1819

2015
[22]

Scalable recommendation with hierarchical poisson factorization

P. Gopalan, J. M. Hofman, and D. M. Blei, “Scalable recommendation with hierarchical poisson factorization.” in UAI, 2015, pp. 326–335

2015
[23]

Probabilistic matrix factorization,

A. Mnih and R. R. Salakhutdinov, “Probabilistic matrix factorization,” in Advances in neural information processing systems , 2008, pp. 1257– 1264

2008
[24]

Factorization meets the neighborhood: a multifaceted col- laborative ﬁltering model,

Y . Koren, “Factorization meets the neighborhood: a multifaceted col- laborative ﬁltering model,” in Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining . ACM, 2008, pp. 426–434

2008
[25]

A scalable collaborative ﬁltering framework based on co-clustering,

T. George and S. Merugu, “A scalable collaborative ﬁltering framework based on co-clustering,” in Data Mining, Fifth IEEE international conference on. IEEE, 2005, pp. 4–pp

2005
[26]

Factorization machines,

S. Rendle, “Factorization machines,” in ICDM. IEEE, 2010, pp. 995– 1000

2010
[27]

Factorization machines with libfm,

——, “Factorization machines with libfm,” TIST, vol. 3, no. 3, p. 57, 2012

2012
[28]

From ranknet to lambdarank to lambdamart: An overview,

C. J. Burges, “From ranknet to lambdarank to lambdamart: An overview,”Learning, vol. 11, no. 23-581, p. 81, 2010

2010
[29]

On the local optimality of lambdarank,

P. Donmez, K. M. Svore, and C. J. Burges, “On the local optimality of lambdarank,” in SIGIR. ACM, 2009, pp. 460–467

2009
[30]

Learning to rank using gradient descent,

C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender, “Learning to rank using gradient descent,” in ICML. ACM, 2005, pp. 89–96

2005
[31]

Repre- sentation learning for very short texts using weighted word embedding aggregation,

C. De Boom, S. Van Canneyt, T. Demeester, and B. Dhoedt, “Repre- sentation learning for very short texts using weighted word embedding aggregation,” Pattern Recognition Letters, vol. 80, pp. 150–156, 2016

2016
[32]

Visualizing data using t-sne,

L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,” JMLR, vol. 9, no. Nov, pp. 2579–2605, 2008

2008
[33]

Celma, Music Recommendation and Discovery in the Long Tail

O. Celma, Music Recommendation and Discovery in the Long Tail . Springer, 2010

2010
[34]

Fma: A dataset for music analysis,

M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “Fma: A dataset for music analysis,” in 18th International Society for Music Information Retrieval Conference , 2017

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Item-based collabo- rative ﬁltering recommendation algorithms,

B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-based collabo- rative ﬁltering recommendation algorithms,” in Proceedings of the 10th International Conference on World Wide Web , ser. WWW ’01. New York, NY , USA: ACM, 2001, pp. 285–295

work page 2001

[2] [2]

A survey of collaborative ﬁltering techniques,

X. Su and T. M. Khoshgoftaar, “A survey of collaborative ﬁltering techniques,” Adv. in Artif. Intell. , vol. 2009, pp. 4:2–4:2, Jan. 2009

work page 2009

[3] [3]

Factorizing personalized markov chains for next-basket recommendation,

S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme, “Factorizing personalized markov chains for next-basket recommendation,” in WWW, 2010, pp. 811–820

work page 2010

[4] [4]

P. Lops, M. de Gemmis, and G. Semeraro, Content-based Recommender Systems: State of the Art and Trends , 2011, p. 73

work page 2011

[5] [5]

How do users describe their information need: Query recommendation based on snippet click model,

Y . Liu, J. Miao, M. Zhang, S. Ma, and L. Ru, “How do users describe their information need: Query recommendation based on snippet click model,” Expert Systems with Applications , vol. 38, no. 11, pp. 13 847 – 13 856, 2011

work page 2011

[6] [6]

Who, where, when, and what: A nonparametric bayesian approach to context-aware recommendation and search for twitter users,

Q. Yuan, G. Cong, K. Zhao, Z. Ma, and A. Sun, “Who, where, when, and what: A nonparametric bayesian approach to context-aware recommendation and search for twitter users,” ACM Trans. Inf. Syst. , vol. 33, no. 1, pp. 2:1–2:33, Feb. 2015

work page 2015

[7] [7]

Social recommendation: a review,

J. Tang, X. Hu, and H. Liu, “Social recommendation: a review,” Social Network Analysis and Mining , vol. 3, no. 4, pp. 1113–1133, 1 2013

work page 2013

[8] [8]

Collaborative ﬁltering with temporal dynamics,

Y . Koren, “Collaborative ﬁltering with temporal dynamics,” Commun. ACM, vol. 53, no. 4, pp. 89–97, Apr. 2010

work page 2010